Improve factor analysis notebook (#612)

* Fix math and code for LS estimate of F

* Clean up

- Use the matrix multiplication operator "@"
- Use \Psi for cov matrix and \sigma for its std

* Spell out some steps in the derivation of amortized inference

* Fix typo

* Use a Bayesian calculation for resampling F

* Add comparison to original data

Also add short sampling of amortized model, to show that it works
but is slow.

* Delete F visualisation

* Allow for negative elements in W

Positive-only works for the particular data we have here, where
the latent factors are defined as positive. But it makes it
impossible to fit models without this constraint.

* Minor clean up

* Add myself to author lists